ip: use rb trees for IP frag queue.

Similar to TCP OOO RX queue, it makes sense to use rb trees to store
IP fragments, so that OOO fragments are inserted faster.

Tested:

- a follow-up patch contains a rather comprehensive ip defrag
  self-test (functional)
- ran neper `udp_stream -c -H <host> -F 100 -l 300 -T 20`:
    netstat --statistics
    Ip:
        282078937 total packets received
        0 forwarded
        0 incoming packets discarded
        946760 incoming packets delivered
        18743456 requests sent out
        101 fragments dropped after timeout
        282077129 reassemblies required
        944952 packets reassembled ok
        262734239 packet reassembles failed
   (The numbers/stats above are somewhat better re:
    reassemblies vs a kernel without this patchset. More
    comprehensive performance testing TBD).
Reported-by: Jann Horn <jannh@google.com>
Reported-by: Juha-Matti Tilli <juha-matti.tilli@iki.fi>
Suggested-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Peter Oskolkov <posk@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>

include/linux/skbuff.h

@@ -676,13 +676,16 @@ struct sk_buff {
 				 * UDP receive path is one user.
 				 */
 				unsigned long		dev_scratch;
-				int			ip_defrag_offset;
 			};
 		};
-		struct rb_node		rbnode; /* used in netem & tcp stack */
+		struct rb_node		rbnode; /* used in netem, ip4 defrag, and tcp stack */
 		struct list_head	list;
 	};
+
+	union {
 		struct sock		*sk;
+		int			ip_defrag_offset;
+	};
 
 	union {
 		ktime_t		tstamp;

include/net/inet_frag.h

@@ -75,7 +75,8 @@ struct inet_frag_queue {
 	struct timer_list	timer;
 	spinlock_t		lock;
 	refcount_t		refcnt;
-	struct sk_buff		*fragments;
+	struct sk_buff		*fragments;  /* Used in IPv6. */
+	struct rb_root		rb_fragments; /* Used in IPv4. */
 	struct sk_buff		*fragments_tail;
 	ktime_t			stamp;
 	int			len;

net/ipv4/inet_fragment.c

@@ -137,12 +137,16 @@ void inet_frag_destroy(struct inet_frag_queue *q)
 	fp = q->fragments;
 	nf = q->net;
 	f = nf->f;
-	while (fp) {
+	if (fp) {
+		do {
 			struct sk_buff *xp = fp->next;
 
 			sum_truesize += fp->truesize;
 			kfree_skb(fp);
 			fp = xp;
+		} while (fp);
+	} else {
+		sum_truesize = skb_rbtree_purge(&q->rb_fragments);
 	}
 	sum = sum_truesize + f->qsize;
 

net/ipv4/ip_fragment.c

@@ -136,7 +136,7 @@ static void ip_expire(struct timer_list *t)
 {
 	struct inet_frag_queue *frag = from_timer(frag, t, timer);
 	const struct iphdr *iph;
-	struct sk_buff *head;
+	struct sk_buff *head = NULL;
 	struct net *net;
 	struct ipq *qp;
 	int err;
@@ -152,14 +152,31 @@ static void ip_expire(struct timer_list *t)
 
 	ipq_kill(qp);
 	__IP_INC_STATS(net, IPSTATS_MIB_REASMFAILS);
-
-	head = qp->q.fragments;
-
 	__IP_INC_STATS(net, IPSTATS_MIB_REASMTIMEOUT);
 
-	if (!(qp->q.flags & INET_FRAG_FIRST_IN) || !head)
+	if (!qp->q.flags & INET_FRAG_FIRST_IN)
 		goto out;
 
+	/* sk_buff::dev and sk_buff::rbnode are unionized. So we
+	 * pull the head out of the tree in order to be able to
+	 * deal with head->dev.
+	 */
+	if (qp->q.fragments) {
+		head = qp->q.fragments;
+		qp->q.fragments = head->next;
+	} else {
+		head = skb_rb_first(&qp->q.rb_fragments);
+		if (!head)
+			goto out;
+		rb_erase(&head->rbnode, &qp->q.rb_fragments);
+		memset(&head->rbnode, 0, sizeof(head->rbnode));
+		barrier();
+	}
+	if (head == qp->q.fragments_tail)
+		qp->q.fragments_tail = NULL;
+
+	sub_frag_mem_limit(qp->q.net, head->truesize);
+
 	head->dev = dev_get_by_index_rcu(net, qp->iif);
 	if (!head->dev)
 		goto out;
@@ -179,16 +196,16 @@ static void ip_expire(struct timer_list *t)
 	    (skb_rtable(head)->rt_type != RTN_LOCAL))
 		goto out;
 
-	skb_get(head);
 	spin_unlock(&qp->q.lock);
 	icmp_send(head, ICMP_TIME_EXCEEDED, ICMP_EXC_FRAGTIME, 0);
-	kfree_skb(head);
 	goto out_rcu_unlock;
 
 out:
 	spin_unlock(&qp->q.lock);
 out_rcu_unlock:
 	rcu_read_unlock();
+	if (head)
+		kfree_skb(head);
 	ipq_put(qp);
 }
 
@@ -231,7 +248,7 @@ static int ip_frag_too_far(struct ipq *qp)
 	end = atomic_inc_return(&peer->rid);
 	qp->rid = end;
 
-	rc = qp->q.fragments && (end - start) > max;
+	rc = qp->q.fragments_tail && (end - start) > max;
 
 	if (rc) {
 		struct net *net;
@@ -245,7 +262,6 @@ static int ip_frag_too_far(struct ipq *qp)
 
 static int ip_frag_reinit(struct ipq *qp)
 {
-	struct sk_buff *fp;
 	unsigned int sum_truesize = 0;
 
 	if (!mod_timer(&qp->q.timer, jiffies + qp->q.net->timeout)) {
@@ -253,20 +269,14 @@ static int ip_frag_reinit(struct ipq *qp)
 		return -ETIMEDOUT;
 	}
 
-	fp = qp->q.fragments;
-	do {
-		struct sk_buff *xp = fp->next;
-
-		sum_truesize += fp->truesize;
-		kfree_skb(fp);
-		fp = xp;
-	} while (fp);
+	sum_truesize = skb_rbtree_purge(&qp->q.rb_fragments);
 	sub_frag_mem_limit(qp->q.net, sum_truesize);
 
 	qp->q.flags = 0;
 	qp->q.len = 0;
 	qp->q.meat = 0;
 	qp->q.fragments = NULL;
+	qp->q.rb_fragments = RB_ROOT;
 	qp->q.fragments_tail = NULL;
 	qp->iif = 0;
 	qp->ecn = 0;
@@ -278,7 +288,8 @@ static int ip_frag_reinit(struct ipq *qp)
 static int ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
 {
 	struct net *net = container_of(qp->q.net, struct net, ipv4.frags);
-	struct sk_buff *prev, *next;
+	struct rb_node **rbn, *parent;
+	struct sk_buff *skb1;
 	struct net_device *dev;
 	unsigned int fragsize;
 	int flags, offset;
@@ -341,58 +352,58 @@ static int ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
 	if (err)
 		goto err;
 
-	/* Find out which fragments are in front and at the back of us
-	 * in the chain of fragments so far.  We must know where to put
-	 * this fragment, right?
-	 */
-	prev = qp->q.fragments_tail;
-	if (!prev || prev->ip_defrag_offset < offset) {
-		next = NULL;
-		goto found;
-	}
-	prev = NULL;
-	for (next = qp->q.fragments; next != NULL; next = next->next) {
-		if (next->ip_defrag_offset >= offset)
-			break;	/* bingo! */
-		prev = next;
-	}
+	/* Note : skb->rbnode and skb->dev share the same location. */
+	dev = skb->dev;
+	/* Makes sure compiler wont do silly aliasing games */
+	barrier();
 
-found:
 	/* RFC5722, Section 4, amended by Errata ID : 3089
 	 *                          When reassembling an IPv6 datagram, if
 	 *   one or more its constituent fragments is determined to be an
 	 *   overlapping fragment, the entire datagram (and any constituent
 	 *   fragments) MUST be silently discarded.
 	 *
-	 * We do the same here for IPv4.
+	 * We do the same here for IPv4 (and increment an snmp counter).
 	 */
 
-	/* Is there an overlap with the previous fragment? */
-	if (prev &&
-	    (prev->ip_defrag_offset + prev->len) > offset)
+	/* Find out where to put this fragment.  */
+	skb1 = qp->q.fragments_tail;
+	if (!skb1) {
+		/* This is the first fragment we've received. */
+		rb_link_node(&skb->rbnode, NULL, &qp->q.rb_fragments.rb_node);
+		qp->q.fragments_tail = skb;
+	} else if ((skb1->ip_defrag_offset + skb1->len) < end) {
+		/* This is the common/special case: skb goes to the end. */
+		/* Detect and discard overlaps. */
+		if (offset < (skb1->ip_defrag_offset + skb1->len))
 			goto discard_qp;
-
-	/* Is there an overlap with the next fragment? */
-	if (next && next->ip_defrag_offset < end)
+		/* Insert after skb1. */
+		rb_link_node(&skb->rbnode, &skb1->rbnode, &skb1->rbnode.rb_right);
+		qp->q.fragments_tail = skb;
+	} else {
+		/* Binary search. Note that skb can become the first fragment, but
+		 * not the last (covered above). */
+		rbn = &qp->q.rb_fragments.rb_node;
+		do {
+			parent = *rbn;
+			skb1 = rb_to_skb(parent);
+			if (end <= skb1->ip_defrag_offset)
+				rbn = &parent->rb_left;
+			else if (offset >= skb1->ip_defrag_offset + skb1->len)
+				rbn = &parent->rb_right;
+			else /* Found an overlap with skb1. */
 				goto discard_qp;
+		} while (*rbn);
+		/* Here we have parent properly set, and rbn pointing to
+		 * one of its NULL left/right children. Insert skb. */
+		rb_link_node(&skb->rbnode, parent, rbn);
+	}
+	rb_insert_color(&skb->rbnode, &qp->q.rb_fragments);
 
-	/* Note : skb->ip_defrag_offset and skb->dev share the same location */
-	dev = skb->dev;
 	if (dev)
 		qp->iif = dev->ifindex;
-	/* Makes sure compiler wont do silly aliasing games */
-	barrier();
 	skb->ip_defrag_offset = offset;
 
-	/* Insert this fragment in the chain of fragments. */
-	skb->next = next;
-	if (!next)
-		qp->q.fragments_tail = skb;
-	if (prev)
-		prev->next = skb;
-	else
-		qp->q.fragments = skb;
-
 	qp->q.stamp = skb->tstamp;
 	qp->q.meat += skb->len;
 	qp->ecn |= ecn;
@@ -414,7 +425,7 @@ static int ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
 		unsigned long orefdst = skb->_skb_refdst;
 
 		skb->_skb_refdst = 0UL;
-		err = ip_frag_reasm(qp, prev, dev);
+		err = ip_frag_reasm(qp, skb, dev);
 		skb->_skb_refdst = orefdst;
 		return err;
 	}
@@ -431,15 +442,15 @@ static int ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
 	return err;
 }
 
-
 /* Build a new IP datagram from all its fragments. */
-
-static int ip_frag_reasm(struct ipq *qp, struct sk_buff *prev,
+static int ip_frag_reasm(struct ipq *qp, struct sk_buff *skb,
 			 struct net_device *dev)
 {
 	struct net *net = container_of(qp->q.net, struct net, ipv4.frags);
 	struct iphdr *iph;
-	struct sk_buff *fp, *head = qp->q.fragments;
+	struct sk_buff *fp, *head = skb_rb_first(&qp->q.rb_fragments);
+	struct sk_buff **nextp; /* To build frag_list. */
+	struct rb_node *rbn;
 	int len;
 	int ihlen;
 	int err;
@@ -453,25 +464,20 @@ static int ip_frag_reasm(struct ipq *qp, struct sk_buff *prev,
 		goto out_fail;
 	}
 	/* Make the one we just received the head. */
-	if (prev) {
-		head = prev->next;
-		fp = skb_clone(head, GFP_ATOMIC);
+	if (head != skb) {
+		fp = skb_clone(skb, GFP_ATOMIC);
 		if (!fp)
 			goto out_nomem;
-
-		fp->next = head->next;
-		if (!fp->next)
+		rb_replace_node(&skb->rbnode, &fp->rbnode, &qp->q.rb_fragments);
+		if (qp->q.fragments_tail == skb)
 			qp->q.fragments_tail = fp;
-		prev->next = fp;
-
-		skb_morph(head, qp->q.fragments);
-		head->next = qp->q.fragments->next;
-
-		consume_skb(qp->q.fragments);
-		qp->q.fragments = head;
+		skb_morph(skb, head);
+		rb_replace_node(&head->rbnode, &skb->rbnode,
+				&qp->q.rb_fragments);
+		consume_skb(head);
+		head = skb;
 	}
 
-	WARN_ON(!head);
 	WARN_ON(head->ip_defrag_offset != 0);
 
 	/* Allocate a new buffer for the datagram. */
@@ -496,24 +502,35 @@ static int ip_frag_reasm(struct ipq *qp, struct sk_buff *prev,
 		clone = alloc_skb(0, GFP_ATOMIC);
 		if (!clone)
 			goto out_nomem;
-		clone->next = head->next;
-		head->next = clone;
 		skb_shinfo(clone)->frag_list = skb_shinfo(head)->frag_list;
 		skb_frag_list_init(head);
 		for (i = 0; i < skb_shinfo(head)->nr_frags; i++)
 			plen += skb_frag_size(&skb_shinfo(head)->frags[i]);
 		clone->len = clone->data_len = head->data_len - plen;
-		head->data_len -= clone->len;
-		head->len -= clone->len;
+		skb->truesize += clone->truesize;
 		clone->csum = 0;
 		clone->ip_summed = head->ip_summed;
 		add_frag_mem_limit(qp->q.net, clone->truesize);
+		skb_shinfo(head)->frag_list = clone;
+		nextp = &clone->next;
+	} else {
+		nextp = &skb_shinfo(head)->frag_list;
 	}
 
-	skb_shinfo(head)->frag_list = head->next;
 	skb_push(head, head->data - skb_network_header(head));
 
-	for (fp=head->next; fp; fp = fp->next) {
+	/* Traverse the tree in order, to build frag_list. */
+	rbn = rb_next(&head->rbnode);
+	rb_erase(&head->rbnode, &qp->q.rb_fragments);
+	while (rbn) {
+		struct rb_node *rbnext = rb_next(rbn);
+		fp = rb_to_skb(rbn);
+		rb_erase(rbn, &qp->q.rb_fragments);
+		rbn = rbnext;
+		*nextp = fp;
+		nextp = &fp->next;
+		fp->prev = NULL;
+		memset(&fp->rbnode, 0, sizeof(fp->rbnode));
 		head->data_len += fp->len;
 		head->len += fp->len;
 		if (head->ip_summed != fp->ip_summed)
@@ -524,7 +541,9 @@ static int ip_frag_reasm(struct ipq *qp, struct sk_buff *prev,
 	}
 	sub_frag_mem_limit(qp->q.net, head->truesize);
 
+	*nextp = NULL;
 	head->next = NULL;
+	head->prev = NULL;
 	head->dev = dev;
 	head->tstamp = qp->q.stamp;
 	IPCB(head)->frag_max_size = max(qp->max_df_size, qp->q.max_size);
@@ -552,6 +571,7 @@ static int ip_frag_reasm(struct ipq *qp, struct sk_buff *prev,
 
 	__IP_INC_STATS(net, IPSTATS_MIB_REASMOKS);
 	qp->q.fragments = NULL;
+	qp->q.rb_fragments = RB_ROOT;
 	qp->q.fragments_tail = NULL;
 	return 0;
 

net/ipv6/netfilter/nf_conntrack_reasm.c

@@ -463,6 +463,7 @@ nf_ct_frag6_reasm(struct frag_queue *fq, struct sk_buff *prev,  struct net_devic
 					  head->csum);
 
 	fq->q.fragments = NULL;
+	fq->q.rb_fragments = RB_ROOT;
 	fq->q.fragments_tail = NULL;
 
 	return true;

net/ipv6/reassembly.c

@@ -405,6 +405,7 @@ static int ip6_frag_reasm(struct frag_queue *fq, struct sk_buff *prev,
 	__IP6_INC_STATS(net, __in6_dev_get(dev), IPSTATS_MIB_REASMOKS);
 	rcu_read_unlock();
 	fq->q.fragments = NULL;
+	fq->q.rb_fragments = RB_ROOT;
 	fq->q.fragments_tail = NULL;
 	return 1;